AI
New Training Method Boosts Chinese Math Reasoning in Open-Source AI Models
Researchers develop a specialized technique to dramatically improve how large language models handle mathematical problems written in Chinese.
A team of researchers has introduced a novel approach to strengthen Chinese mathematical reasoning in large language models, addressing a significant gap in multilingual AI capabilities. Using the open-source Llama-2-13B model as their test case, they achieved a substantial 36-percentage-point improvement in Chinese math problem accuracy.
The Challenge of Multilingual Capability#
While mainstream AI systems like ChatGPT have demonstrated impressive abilities across multiple languages, many smaller or specialized models struggle with specific language-and-task combinations. Extending these systems to handle new languages typically demands either expensive retraining from scratch or sacrificing performance on existing languages. The researchers tackled this by focusing on how to rapidly enhance a model's competency in a targeted domain—Chinese mathematics—without compromising its existing strengths.
New Training Techniques#
The solution centers on two complementary methods:
-
Similarity-constrained RRHF (SimRRHF): This refinement of Reward Ranking and Ranking Feedback training adds a similarity loss term that guides the model's learning direction, preventing it from straying too far from its original capabilities while absorbing new knowledge.
-
Incorrect Data Relearning (IDRL): Designed to help models master particularly difficult concepts, this technique refocuses training on errors the model makes during reasoning tasks, strengthening its ability to learn from mistakes.
Results and Broader Impact#
The combined approach, called Olapa-MCoT (Mathematical Chain of Thought), produced striking results: Chinese mathematical reasoning accuracy reached 50%, up from just 14% in the base model. The improvements extended beyond the target language—English mathematical reasoning also gained roughly 4 percentage points. The researchers emphasize that their method is adaptable to other languages and model architectures, offering a template for quickly expanding AI capabilities in underserved linguistic domains.
